Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
25
25
image
imagewidth (px)
1.92k
1.92k
rgba
imagewidth (px)
1.92k
1.92k
mask
imagewidth (px)
1.92k
1.92k
aaardnntevvzd_tmptu009ksn
aaaxmlmrpcivg_tmpj3tw6s1f
aabdngrwjaftg_tmptkkgckkv
aadexkvperlur_tmp8s3ta1rx
aadjztmwavuij_tmpsh9ojjut
aafajtwjtyinf_tmp8z4j9vhb
aahhwjlvdxzxd_tmpbvwiz_hz
aajjvsrtwobbq_tmpuormai0j
aalvkywqwglcz_tmps7n75g7b
aapwskcgrdjjr_tmpkijnqmhy
aaqejoqxcczkd_tmp0uhvpqyw
aaqwisqmpmhzm_tmp7sj_r9k3
aawghtezexdjp_tmprk06iwfp
aaxrqsnwzzezz_tmp6eft8o8m
abirxjtimsuzg_tmpc6exufo4
abmplzcdrtkvu_tmp90p474_m
aboxeevjpoydt_tmpf0477nwd
abpdvgobdvaxe_tmphjcee8vo
absdmzszxohqr_tmpnstrwv8a
abtjnmpchmctm_tmpfoxzqj3r
abtzkycnjuqzt_tmpv_s7gl0b
abygiaweoomra_tmp2r3cgy4j
abyhwpjeuyevb_tmp1n38ulkk
abytydvkngrjf_tmpsjifuqym
abzjmtunafuhx_tmpa_2p8kit
acaehzqvfjurc_tmp8v_kr11k
accrthyxijhhq_tmp2tgx1ueg
acepvpwymbwpm_tmpyjcrscke
aciugfxmkwvsa_tmp785m2b2d
ackfqoqudpafm_tmpsvga9nbe
ackssleenwulx_tmp8v4wwuq_
aclbfpkxqxhxu_tmpszkxo2dw
acminphjcschx_tmp3cn054h3
acsubfvltfixi_tmpkkfhhxzz
actgnsnztvcdu_tmpoub30akm
acvtcsarrfxcr_tmp5l82hnpq
acwaepixfxypf_tmpxsqfe4tx
acwcypvuqebbm_tmpccb6uf8s
acwjnrmbnprgs_tmptivmkaej
acwtpkvsmozuk_tmpxyguz1yw
acwwqsizdbpth_tmpa8bt0lsp
acxnxkpqcdnyo_tmpikf16bwp
acyjexltoyupc_tmpfggytj_i
adcoyjxgttjef_tmp3tb7atsz
adcqiixdteaqn_tmphipwvnvw
adfvywbtnjoct_tmpqe5srxns
adjqjmadhwxzb_tmpdbt8zdya
adjuafbvziwhv_tmpgi2sulmh
adnilhsrrakud_tmprxy1gmzj
adqyfsvhqeown_tmpfw9uxgi9
adtrryiejjwbm_tmpznil3rzg
adtzgjctiafbd_tmp62b7vk5v
aduakfesvlhup_tmpfl1n1j5b
adxgdsxssxxjm_tmpzm8uodn3
adyfrxivroixo_tmprzstr6ao
adzdcejjnqexl_tmp5wy7weg0
adzeiuxdhhcuz_tmpbcyn9oj_
adzhcdlfmrzqj_tmprc6mj08y
aeafbuoajnudl_tmp2pucsx3l
aefjsmfjwzmkh_tmpg_tqhh8y
aejccsdgadpyr_tmpzweqvf_h
aenhxlxzpfjkn_tmp36mbunev
aenqbyanavnbq_tmpf4f6_mn6
aeoajswhprend_tmptfn30z9p
aepgneoqllvhg_tmp_gyllnj2
aerljxvpxgchi_tmp4_8y1uzm
aespvuzitfthd_tmpegnk7udo
aetmjopowknvo_tmpvj46leck
aevakdayonhir_tmptogot68g
aeypemmzhdhhm_tmpirt74pi3
aeyvhiyqqcgot_tmpfq5n2mvh
afbfgbqwvxeho_tmph2a29mcy
afbhvmkvwbsij_tmpalwfvr3s
afevwoexdavgq_tmp0xvf41zs
afiblpkwjgtkb_tmpteyoz8pf
afifiiitaltcj_tmp4s4c0btf
afjktszhnkpjm_tmp1q8lwpkk
afjntvuigomsc_tmpki4zajhj
aflsepehccvjy_tmph2euz82r
afmuhqqrssfzn_tmpv45ge0bs
afqilqnrffwlc_tmpx4e37_sq
afszvaptcsutw_tmpmqd0numf
aftglbxpqyzyq_tmpyzkjo318
afttccwfeemcd_tmprx32xiyy
afvdwmidqzahx_tmpt27e5l4g
afwdegddelgfn_tmpehhr_pf9
afwkffevputbz_tmpc03l2u2f
afxkbezpjlcuj_tmp5m4nxoy0
afzyjmsudtgbv_tmp2jhcch4p
agakdpomcclsd_tmpp3vna19j
agapljbrlhapp_tmpozutekvt
agchwuzyhftps_tmp78047ek2
aghldfqeckvwl_tmp7iontsei
agkthhizarfxk_tmpxgm6p7sr
agowfniwcxzmw_tmp1pbbdbc6
agtvupxdxbaes_tmplezc8285
agtzmvfqtsjuz_tmppevn2nm9
aguylobwmadld_tmpf5f5rhae
agzjorwjlmksw_tmpu3vz0hg2
ahadxnaqqbpor_tmp53ms95jk
End of preview. Expand in Data Studio

Humans RGBA proxy pairs (historical revision)

11,665 synthetic human images, each with its RGBA pair and its 8-bit alpha mask. There are 9,116 train and 2,549 validation rows. The splits are the upstream training and validation folders.

from datasets import load_dataset

ds = load_dataset("nobg/humans-rgba")
ds["train"]        # 9,116 rows
ds["validation"]   # 2,549 rows
row = ds["train"][0]
row["id"]      # 'aaardnntevvzd_tmptu009ksn', the upstream file stem
row["image"]   # PIL RGB, the original upstream PNG
row["rgba"]    # PIL RGBA, the same RGB with the upstream gt as alpha
row["mask"]    # PIL L, the alpha channel of rgba

Not an isolated-foreground RGBA ground truth. image contains the composited background, including under translucent edges and fully transparent pixels. rgba copies that composited RGB unchanged. It does not recover foreground color. The original masks are segmentation and matting targets generated from synthetic transparent humans. They are not necessarily hand-annotated real photos.

Why this mirror exists

The previous revision of this repo stored the pairs as 365 custom TAR shards with <split>/<id>/source.png and <split>/<id>/target.png members, plus a manifest.json. The Hugging Face dataset viewer could not read that layout, and load_dataset needed custom code. This revision stores the same pairs as parquet with datasets Image features. The viewer renders it, and load_dataset works with no extra code.

It also adds a mask column. That matches the image and mask columns of the other nobg sets, so the set can train a mask model as well as an RGBA model.

Columns

column type content
id string Upstream file stem, shared by im/<id>.png and gt/<id>.png
image Image (RGB PNG) The upstream im PNG. Bytes are unmodified.
rgba Image (RGBA PNG) image with the upstream gt PNG as alpha. Bytes are unmodified from the previous revision.
mask Image (L PNG) The alpha channel of rgba, saved as an 8-bit grayscale PNG.

Processing

  • Source: the previous revision of nobg/humans-rgba, 365 TAR shards and manifest.json.
  • Every TAR was checked against its SHA-256 in manifest.json. Member names had to match the manifest exactly. Each TAR was read in memory and never extracted to disk.
  • image holds the source.png bytes and rgba holds the target.png bytes. Neither was decoded and re-encoded. The SHA-256 of both columns matches the manifest on all 11,665 rows.
  • mask is the A band of rgba, written with Pillow as a PNG in mode L. Decoding it gives exactly the alpha of rgba on every row. Soft alpha values are kept, not thresholded.
  • On every row, image, rgba and mask have the same size, and the RGB channels of rgba equal image pixel for pixel.
  • Upstream training is renamed to train. validation keeps its name. The assignment of rows to splits is unchanged. No id appears in both splits.
  • Rows are ordered by upstream id within each split. Shards are about 500 MB: 166 for train and 47 for validation, 111.9 GB in total.
  • No background or transparent-human originals are included.

The original pairs were built by the earlier revision as follows. source.png is the upstream im PNG. target.png is that RGB with the upstream 8-bit grayscale gt PNG attached as alpha.

Provenance

Derived only from schirrmacher/humans, revision aa149eb852ecd5845bdc7a9324eb41ed4f6eef46 (schirrmacher/humans@aa149eb852ecd5845bdc7a9324eb41ed4f6eef46), folders training/im, training/gt, validation/im and validation/gt. Creator: Marvin Schirrmacher.

Historical snapshot, not the current advertised release. The pinned revision predates a later dataset update. Its 9,116 and 2,549 pair counts are verified from its file tree. This is not the current release of about 7k images. The original train and validation assignment is retained unchanged.

Licensing

The pinned dataset card declares Apache-2.0. Please retain this provenance, the original attribution, and applicable Apache-2.0 notices. The source card describes humans and backgrounds generated with LayerDiffuse and Stable Diffusion. Individual upstream image rights have not been independently verified.

Citation

Cite the original dataset:

Marvin Schirrmacher, Human Segmentation Dataset, https://huggingface.co/datasets/schirrmacher/humans

@misc{schirrmacher_humans,
  author = {Schirrmacher, Marvin},
  title = {Human Segmentation Dataset},
  howpublished = {\url{https://huggingface.co/datasets/schirrmacher/humans}}
}
Downloads last month
112